When you use in access log handler multiple patterns for response time, mainly visible when using one for microseconds and second for nanoseconds, the times differ as the time is not counted based on when the response really ended but based on when the io.undertow.attribute.ResponseTimeAttribute#readAttribute is called, which for each of the ResponseTimeAttributes slightly differs.

This leads to slight differences in the reported times.

I belive the reported time should be the same (with exception of converting to particular unit and related loss of precision) as the response took the same time.

The solution might be to storing also response end time in the exchange and counting the response time based on it.